Steel design based on a large language model

The success of artificial intelligence (AI) in materials research heavily relies on the integrity of structured data and the construction of precise descriptors. In this study, we present an end-to-end pipeline from materials text to properties for steels based on a large language model. The objective is to enable quantitative predictions of properties with high-accuracy and explore new steels. The pipeline includes a materials language encoder, named SteelBERT, and a multimodal deep learning framework that maps the composition and text sequence of complex fabrication processes to mechanical properties. We demonstrate high accuracy on mechanical properties, including yield strength (YS), ultimate tensile strength (UTS), and elongation (EL) by predicting determination coefficients (R²) reaching 78.17 % ( ± 3.40 %), 82.56 % ( ± 1.96 %), and 81.44 % ( ± 2.98 %) respectively. Further, through an additional fine-tuning strategy for the design of specific steels with small datasets, we show how the performance can be refined. With only 64 experimental samples of 15Cr austenitic stainless steels, we obtain an optimized model with R² of 89.85 % ( ± 6.17 %), 88.34 % ( ± 5.95 %) and 87.24 % ( ± 5.15 %) for YS, UTS and EL, that requires the user to input composition and text sequence for processing and which outputs mechanical properties. The model efficiently optimizes the text sequence for the fabrication process by suggesting a secondary round of cold rolling and tempering to yield an exceptional YS of 960 MPa, UTS of 1138 MPa, and EL of 32.5 %, exceeding those of reported 15Cr austenitic stainless steels.